Starting to dive into gpt4. One thing off the top, is how much better it is at following a technical conversation involving mathematics. I will be looking for it to remember defintions make during a conversation and consistently use the nomenclature we define. For me the biggest interest right now is exploring complex algorithms by teaching it how to do them and then exploring the properties of this algorithms. To do this effectively it has to be able to remember in context what we defined. Being able to have it digest specific web pages is a feature I hope will be coming soon. It has the potential to become more useful than a beginning graduate student as a research assistant.
Welcome to the community!
I would caution you against relying on gpt-4 alone for such tasks. It can do a lot of stuff, but there are odd cases where it will fail, with confidence - poisoning the thread.
Do you have any thoughts on how to mitigate that?
Affirmed. I catch it making errors. Giving it the ability to assess its own error rate would be good. The sky’s the limit, if what is made available can be counted on to understand parameters we define to bind it. Then I can use it to do all the heavy lifting and research tasks. It becomes more like a piece of computational clay that we can mold and statistically sample. I think for basic use, thinking of it as an apprentice and not a journeyman, if it could access webpages and quickly and accurately digest the contents, and… more than that, since it can write code, actually learn functionality from websites and codebases how powerful would that be? The need to include graphics in its responses seems big to me. If sora can develop simulations, the teaching assistant aspect is mind boggling, the thought experiment aspect. I’m just starting to think about all the ways this kind of extension of my abilities would benefit me. As an expert system, say, I create one and talk to it about my garden, I want it to be able to gather data about my plot and assimilate what I tell it into a narrative about choices I have for farming. So if it knows my USGS soil types and has a GIS database connection, I could slowly teach it to help me answer questions like, “What might be a good place and time for me to plant Brussels sprouts.” “Help me fine tune a production schedule for my blue tulips.” Of course the cool part is not that it might give me a table,but that it bounces of off my own consciousness conversationally. Time will tell, but I am pretty certain that this is going to be fantastic tool for teachers, researchers, governments, and just everyday people, once it develops further and becomes more widely understood. I’m very positive about it.